H3 video generation: a case library you can actually read
Most AI video showcases show you a finished clip and nothing else. This project takes the opposite approach: it collects 1,274 real MiniMax H3 generations from public sources, keeps the prompt that produced each one, and labels every entry along dimensions that are useful when you are trying to write a prompt yourself — what the shot is doing, how the camera moves, what the output looks like.
You can browse it at the H3 case library. Every entry links back to where it was published. This page explains what is in there, how the labels were produced, and where the limits are.
What the library contains
The current snapshot holds 1,274 cases. Of those, the majority are landscape, short, and single-generation clips — that is simply what people publish.
| Aspect ratio | Landscape 1,060 · Portrait 190 · Square 24 |
|---|
| Duration | 8–16s 1,101 · under 8s 92 · over 16s 81 |
|---|
| Generation mode | Unverified 930 · Text-to-video 164 · Reference-to-video 128 · First/last-frame 49 · Image-to-video (unspecified) 3 |
|---|
| Topic labels | 26 values across short-drama dialogue, product ads, character performance, music and dance, landscapes, fashion, city architecture, animation, martial arts and more |
|---|
| Shot-technique labels | 28 values — close-ups, dolly moves, tracking shots, aerial work, lip-sync, on-screen text, multi-shot switching, first/last-frame anchoring |
|---|
| Visual-style labels | 8 values — cinematic, photorealistic, 2D animation, 3D animation, slow motion, stop-motion collage, black and white |
|---|
| Speech detection | Dialogue detected in 251 entries; 314 have no speech; 704 were not analysed in this round |
|---|
How each entry is labelled
Every case carries the same set of fields, so the library can be filtered and sorted rather than just scrolled:
- Topic, shot technique, visual style — multi-value labels. A clip can be both “product ad” and “close-up”.
- Generation mode — text-to-video, reference-to-video, first/last-frame, or unverified.
- Duration, resolution, aspect ratio — measured from the source file, not from the poster's description.
- Prompt length — how long the prompt behind the clip actually was.
- Record status — whether the metadata has been semantically reviewed, or is still flagged as conflicting.
- Review flags — anything a human needs to look at again.
Labels are produced as candidate tags and then reviewed. Anything not yet confirmed is marked as pending review rather than silently presented as fact — that is why some entries show a “pending review” tag next to a real topic tag.
How to use it
Two ways, depending on what you are trying to learn:
- If you are writing a prompt — filter by shot technique, then read the prompts of the clips whose camera work you want. The interesting signal is not the subject matter, it is how the prompt describes motion, timing and what happens at the start and end of the shot.
- If you are surveying the model — sort by prompt length and look at the long-prompt cases. They show what the model can absorb, and they are where the unusual results live.
Search matches topic labels, shot labels, author, internal ID and title, so you can jump straight to a specific creator or a specific case ID.
What this library does not claim
Being explicit about the limits matters more than sounding confident:
- No quality ranking. Inclusion means a clip was published publicly and could be labelled — not that it is good. Nothing here is a ranking of models, creators or outputs.
- Generation mode is often unknown. 930 of 1,274 entries are marked “unverified”. Publishers rarely state how a clip was produced, and it cannot be reliably inferred from the video.
- Speech detection is partial. It run on a subset, and three entries are flagged as suspected transcription artefacts rather than genuine dialogue.
- Prompts come with the clip, not from the model vendor. Where a prompt was published by the creator, it is reproduced as-is and attributed. Not every case has one.
- Videos are played from their original hosts. Availability depends on the source site; a link that worked last month may not work today.
Data provenance and removal requests
The library is assembled from publicly published H3 outputs. Prompts and videos remain the property of their original authors, and every entry links to its original publication so attribution is one click away. If you are an author and want an entry removed, ask and it will be taken down.
The deployment and setup material for running H3 on cloud GPUs is maintained separately and documented in Chinese; the overview of that work is not published on this page.